@anteros/core 0.0.41 → 0.0.42

package/lib/attachment.ts

@@ -0,0 +1,351 @@
+import path from "path";
+import { AppError } from "./error";
+import { openZip, type ZipArchive } from "./zip";
+import type { AgentAttachment, AgentMediaPart, AgentTextPart } from "../types/agent";
+
+/**
+ * Attachments — images and documents handed to a model with a run.
+ *
+ * An LLM only understands three things: **text**, **images** and — for some
+ * models — **documents** (PDF). So every file is resolved to one of those:
+ *
+ * | Input | What the model receives |
+ * | --- | --- |
+ * | `.md` `.txt` `.json` `.csv` `.tsv` `.yaml` `.xml` `.html` `.log`, code… | the **text**, under a header naming the file |
+ * | `.png` `.jpg` `.jpeg` `.webp` `.gif` | an **image** part (base64) — native vision |
+ * | `.pdf` | a **file** part (base64) — native document input, where the model supports it |
+ * | `.docx` | the **text** extracted from `word/document.xml` (no dependency: the file is a ZIP) |
+ * | `.xlsx` | the **rows**, one CSV block per sheet, from the workbook parts |
+ * | `.doc` (legacy), anything else | refused — `AGENT_FILE_UNSUPPORTED` |
+ *
+ * The extraction of `.docx` / `.xlsx` is a deliberate simplification: it reads
+ * the text content, not the layout, the styles or the embedded objects. Good
+ * enough for a model; never a replacement for a document renderer.
+ */
+
+/** Text formats read as-is — the cheapest and most faithful thing to do. */
+const TEXT_EXTENSIONS = new Set([
+    "md", "markdown", "mdx", "txt", "text", "log",
+    "json", "jsonc", "ndjson", "csv", "tsv",
+    "yaml", "yml", "toml", "ini", "env",
+    "xml", "html", "htm", "svg",
+    "ts", "tsx", "js", "jsx", "mjs", "cjs", "py", "rb", "go", "rs", "java", "php",
+    "sql", "sh", "bash", "zsh", "css", "scss", "vue", "svelte", "graphql", "gql",
+]);
+
+/** Image formats both providers accept natively. */
+const IMAGE_MIME: Record<string, string> = {
+    png: "image/png",
+    jpg: "image/jpeg",
+    jpeg: "image/jpeg",
+    webp: "image/webp",
+    gif: "image/gif",
+};
+
+/** Formats the model can receive as a document part. */
+const DOCUMENT_MIME: Record<string, string> = {
+    pdf: "application/pdf",
+};
+
+/** Extensions worth naming when we refuse a file. */
+const OFFICE_BINARY: Record<string, string> = {
+    doc: "the legacy .doc format is not readable — convert it to .docx or PDF",
+    xls: "the legacy .xls format is not readable — convert it to .xlsx or CSV",
+    ppt: "the legacy .ppt format is not readable — convert it to .pptx or PDF",
+    pptx: "PowerPoint is not extracted — convert it to PDF",
+};
+
+/** One file never exceeds this — a model is not a file server. */
+const MAX_FILE_BYTES = 10 * 1024 * 1024;
+/** Inlined text is capped too: a 40 MB CSV must not become a 40 MB prompt. */
+const MAX_TEXT_CHARS = 200_000;
+
+function extensionOf(name: string): string {
+    return path.extname(name).replace(/^\./, "").toLowerCase();
+}
+
+function unsupported(name: string): AppError {
+    const extension = extensionOf(name);
+    const hint = OFFICE_BINARY[extension];
+    return new AppError(
+        hint
+            ? `Cannot read '${name}': ${hint}`
+            : `Cannot read '${name}': supported attachments are images (png, jpg, webp, gif), PDF, `
+            + `docx, xlsx and text formats (md, txt, json, csv, yaml, code…)`,
+        { status: 400, code: "AGENT_FILE_UNSUPPORTED" },
+    );
+}
+
+function tooLarge(name: string, size: number): AppError {
+    return new AppError(
+        `'${name}' is ${Math.round(size / 1024 / 1024)}MB — the limit is ${MAX_FILE_BYTES / 1024 / 1024}MB`,
+        { status: 413, code: "AGENT_FILE_TOO_LARGE" },
+    );
+}
+
+function toBase64(bytes: Uint8Array): string {
+    return Buffer.from(bytes).toString("base64");
+}
+
+/** Decode the XML entities an Office part can contain. */
+function decodeEntities(text: string): string {
+    return text
+        .replace(/&lt;/g, "<")
+        .replace(/&gt;/g, ">")
+        .replace(/&quot;/g, '"')
+        .replace(/&apos;/g, "'")
+        .replace(/&#(\d+);/g, (_, code) => String.fromCharCode(Number(code)))
+        .replace(/&amp;/g, "&");
+}
+
+function stripTags(xml: string): string {
+    return decodeEntities(xml.replace(/<[^>]*>/g, ""));
+}
+
+// ─── Word ────────────────────────────────────────────────────────────────
+
+/** `word/document.xml` → plain text (paragraphs, tabs and breaks preserved). */
+function docxText(zip: ZipArchive): string {
+    const document = zip.read("word/document.xml");
+    if (!document) throw new AppError("Not a readable .docx (no word/document.xml)", {
+        status: 400, code: "AGENT_FILE_INVALID",
+    });
+
+    const xml = new TextDecoder().decode(document);
+    const text = xml
+        .replace(/<w:tab\b[^>]*\/>/g, "\t")
+        .replace(/<w:br\b[^>]*\/>/g, "\n")
+        .replace(/<\/w:p>/g, "\n")
+        .replace(/<[^>]*>/g, "");
+
+    return decodeEntities(text)
+        .replace(/[ \t]+\n/g, "\n")
+        .replace(/\n{3,}/g, "\n\n")
+        .trim();
+}
+
+// ─── Excel ───────────────────────────────────────────────────────────────
+
+/** `xl/sharedStrings.xml` → the strings the cells refer to by index. */
+function xlsxSharedStrings(zip: ZipArchive): string[] {
+    const shared = zip.read("xl/sharedStrings.xml");
+    if (!shared) return [];
+
+    const xml = new TextDecoder().decode(shared);
+    return [...xml.matchAll(/<si\b[^>]*>([\s\S]*?)<\/si>/g)].map((match) =>
+        [...(match[1] ?? "").matchAll(/<t\b[^>]*>([\s\S]*?)<\/t>/g)]
+            .map((run) => decodeEntities(run[1] ?? ""))
+            .join(""));
+}
+
+/** `B7` → 1 (0-based column index), so gaps in a row are preserved. */
+function columnOf(reference: string): number {
+    const letters = /^([A-Z]+)/.exec(reference)?.[1] ?? "";
+    let index = 0;
+    for (const letter of letters) index = index * 26 + (letter.charCodeAt(0) - 64);
+    return Math.max(0, index - 1);
+}
+
+function csvCell(value: string): string {
+    return /[",\n]/.test(value) ? `"${value.replace(/"/g, '""')}"` : value;
+}
+
+/** One worksheet → CSV rows (values, not formulas — a model reads values). */
+function xlsxSheet(xml: string, shared: string[]): string[] {
+    const rows: string[] = [];
+
+    for (const rowMatch of xml.matchAll(/<row\b[^>]*>([\s\S]*?)<\/row>/g)) {
+        const cells: string[] = [];
+        for (const cellMatch of (rowMatch[1] ?? "").matchAll(/<c\b([^>]*)>([\s\S]*?)<\/c>/g)) {
+            const attributes = cellMatch[1] ?? "";
+            const body = cellMatch[2] ?? "";
+            const type = /\bt="([^"]+)"/.exec(attributes)?.[1];
+            const reference = /\br="([^"]+)"/.exec(attributes)?.[1] ?? "";
+            const position = reference ? columnOf(reference) : cells.length;
+
+            let value = "";
+            if (type === "s") {
+                const index = Number(stripTags(/\<v\>([\s\S]*?)\<\/v\>/.exec(body)?.[1] ?? ""));
+                value = shared[index] ?? "";
+            } else if (type === "inlineStr") {
+                value = [...body.matchAll(/<t\b[^>]*>([\s\S]*?)<\/t>/g)]
+                    .map((run) => decodeEntities(run[1] ?? "")).join("");
+            } else {
+                value = decodeEntities(/\<v\>([\s\S]*?)\<\/v\>/.exec(body)?.[1] ?? "");
+            }
+
+            while (cells.length < position) cells.push("");
+            cells[position] = value;
+        }
+        rows.push(cells.map(csvCell).join(","));
+    }
+
+    return rows;
+}
+
+/** Every worksheet, in tab order, as `### <sheet name>` + CSV rows. */
+function xlsxText(zip: ZipArchive): string {
+    const sheets = zip.names()
+        .filter((name) => /^xl\/worksheets\/sheet\d+\.xml$/.test(name))
+        .sort((a, b) => Number(/(\d+)\.xml$/.exec(a)?.[1]) - Number(/(\d+)\.xml$/.exec(b)?.[1]));
+
+    if (!sheets.length) throw new AppError("Not a readable .xlsx (no worksheet)", {
+        status: 400, code: "AGENT_FILE_INVALID",
+    });
+
+    const shared = xlsxSharedStrings(zip);
+
+    return sheets.map((sheet, index) => {
+        const xml = new TextDecoder().decode(zip.read(sheet) ?? new Uint8Array());
+        return `### Sheet ${index + 1}\n\n${xlsxSheet(xml, shared).join("\n")}`;
+    }).join("\n\n");
+}
+
+// ─── Resolution ──────────────────────────────────────────────────────────
+
+/** What a caller can attach — a path (server-side), a `Blob`/`File`, raw bytes, or base64. */
+export type NormalizedAttachment = {
+    name: string;
+    mimeType: string;
+    bytes: Uint8Array;
+};
+
+/**
+ * Read what the caller attached. A string is a **path** on the server (read
+ * now, at call time) — the SDK sends base64 instead, since a browser has no
+ * filesystem.
+ */
+async function materialize(attachment: AgentAttachment, baseDir: string): Promise<NormalizedAttachment> {
+    if (typeof attachment === "string") {
+        const file = path.resolve(baseDir, attachment);
+        if (!(await Bun.file(file).exists())) {
+            throw new AppError(`Attachment '${attachment}' not found`, { status: 400, code: "AGENT_FILE_NOT_FOUND" });
+        }
+        const bytes = new Uint8Array(await Bun.file(file).arrayBuffer());
+        return { name: path.basename(file), mimeType: "", bytes };
+    }
+
+    if (attachment instanceof Uint8Array) {
+        return { name: "attachment", mimeType: "", bytes: attachment };
+    }
+
+    if (typeof Blob !== "undefined" && attachment instanceof Blob) {
+        const name = (attachment as File).name ?? "attachment";
+        return {
+            name,
+            mimeType: (attachment as File).type ?? "",
+            bytes: new Uint8Array(await attachment.arrayBuffer()),
+        };
+    }
+
+    const explicit = attachment as { name?: string; mimeType?: string; data: string | Uint8Array; encoding?: "base64" | "utf8" };
+    const name = explicit.name ?? "attachment";
+    if (explicit.data instanceof Uint8Array) {
+        return { name, mimeType: explicit.mimeType ?? "", bytes: explicit.data };
+    }
+
+    const encoding = explicit.encoding ?? "base64";
+    return {
+        name,
+        mimeType: explicit.mimeType ?? "",
+        bytes: encoding === "utf8"
+            ? new TextEncoder().encode(explicit.data)
+            : new Uint8Array(Buffer.from(explicit.data, "base64")),
+    };
+}
+
+/**
+ * Turn attachments into **content parts** the providers understand.
+ *
+ * @returns the parts to append to the current user turn (`[]` when there is
+ * nothing to attach). Throws `AGENT_FILE_*` when a file cannot be sent — a
+ * silently dropped attachment is worse than a loud failure.
+ */
+async function resolveAttachments(
+    attachments: AgentAttachment | AgentAttachment[] | undefined,
+    options: { baseDir?: string } = {},
+): Promise<AgentMediaPart[]> {
+    if (!attachments) return [];
+    const list = Array.isArray(attachments) ? attachments : [attachments];
+    if (!list.length) return [];
+
+    const baseDir = options.baseDir ?? process.cwd();
+    const parts: AgentMediaPart[] = [];
+
+    for (const attachment of list) {
+        const file = await materialize(attachment, baseDir);
+        if (file.bytes.byteLength > MAX_FILE_BYTES) throw tooLarge(file.name, file.bytes.byteLength);
+
+        const extension = extensionOf(file.name);
+        const declared = file.mimeType.toLowerCase();
+        const mimeType = declared || IMAGE_MIME[extension] || DOCUMENT_MIME[extension] || "";
+
+        // 1. Images — native vision on both protocols
+        const imageMime = IMAGE_MIME[extension] ?? (declared.startsWith("image/") ? declared : "");
+        if (imageMime) {
+            parts.push({ type: "image", data: toBase64(file.bytes), mimeType: imageMime });
+            continue;
+        }
+
+        // 2/3. PDF — a document part (the provider decides what it does with it)
+        const documentMime = DOCUMENT_MIME[extension] ?? (declared === "application/pdf" ? declared : "");
+        if (documentMime) {
+            parts.push({ type: "file", name: file.name, mimeType: documentMime, data: toBase64(file.bytes) });
+            continue;
+        }
+
+        // 4. Office documents — extracted to text, no dependency
+        if (extension === "docx" || extension === "xlsx") {
+            const text = extension === "docx"
+                ? docxText(openZip(file.bytes))
+                : xlsxText(openZip(file.bytes));
+            parts.push({ type: "text", text: textPart(file.name, text) });
+            continue;
+        }
+
+        // 5. Text formats — read as-is
+        if (TEXT_EXTENSIONS.has(extension) || declared.startsWith("text/") || declared === "application/json") {
+            const text = new TextDecoder().decode(file.bytes);
+            parts.push({ type: "text", text: textPart(file.name, text) });
+            continue;
+        }
+
+        throw unsupported(file.name);
+    }
+
+    return parts;
+}
+
+/** A file's text under a header naming it — the model must know where it comes from. */
+function textPart(name: string, text: string): string {
+    const trimmed = text.trim();
+    if (trimmed.length <= MAX_TEXT_CHARS) return `--- ${name} ---\n${trimmed}`;
+    return `--- ${name} (truncated) ---\n${trimmed.slice(0, MAX_TEXT_CHARS)}`;
+}
+
+/**
+ * What a **memory** keeps of an attachment. Conversations are text: storing a
+ * 5 MB image in every turn would fill the store for nothing, and the model
+ * already read it once. The reply and the question stay verbatim.
+ */
+function attachmentNote(part: AgentMediaPart): string {
+    if (part.type === "image") return "[image]";
+    if (part.type === "file") return `[file ${part.name ?? "document"}]`;
+    const header = /^--- (.+?) ---/.exec(part.text)?.[1];
+    return header ? `[file ${header}]` : "";
+}
+
+/** A part as it should be *stored* — the heavy payloads replaced by a note. */
+function forMemory(part: AgentMediaPart): AgentTextPart {
+    if (part.type === "text") return part;
+    return { type: "text", text: attachmentNote(part) };
+}
+
+export {
+    MAX_FILE_BYTES,
+    MAX_TEXT_CHARS,
+    docxText,
+    forMemory,
+    resolveAttachments,
+    xlsxText,
+};

package/lib/auditResult.ts

@@ -107,6 +107,19 @@ function summarizeAuditResult(action: string, result: any): any {
             return typeof result === 'string' ? result : null
         case 'dropIndexes':
             return result?.nIndexesWas !== undefined ? { nIndexesWas: result.nIndexesWas } : null
+        case 'agent.generate':
+        case 'agent.stream':
+        case 'agent.object':
+            // The model's answer is never stored — only the envelope that proves the
+            // run happened and what it cost (nothing the caller did not already send).
+            return result ? {
+                finishReason: result.finishReason ?? null,
+                usage: result.usage ?? null,
+                tools: Array.isArray(result.toolCalls)
+                    ? result.toolCalls.map((call: any) => call?.name).filter(Boolean)
+                    : [],
+                steps: Array.isArray(result.steps) ? result.steps.length : null,
+            } : null
         default:
             // Anything unknown: a document must never be stored.
             return null

package/lib/define.ts

@@ -159,14 +159,54 @@ function Middleware(config: GlobalMiddlewareConfig): GlobalMiddlewareConfig {
 }
 
 import type { WorkflowDefinition } from "../types/workflow";
+import type { AgentConfig, AgentTool, AgentToolDefinition } from "../types/agent";
 
 export function Workflow<TData = any>(workflow: WorkflowDefinition<TData>): WorkflowDefinition<TData> {
     return {
         ...workflow,
+        enabled: workflow.enabled ?? true,
         _isWorkflow_: true,
     }
 }
 
+/**
+ * Declares an LLM agent — `{tenant.dir}/agents/**\/*.agent.ts`.
+ *
+ * `id` and `description` are **required**: the id identifies it (and looks it up),
+ * the description says what it is for — it is served by `info` and read by humans.
+ *
+ * @example
+ * ```ts
+ * export default define.Agent({
+ *   id: 'weather',
+ *   description: 'Answers weather questions with a live forecast.',
+ *   instructions: 'You are a concise weather assistant.',
+ *   provider: { model: 'gpt-4o-mini', compatible: 'openai' },
+ *   tools: { forecast },
+ * })
+ * ```
+ */
+function Agent(config: AgentConfig & { id: string; description: string }): AgentConfig {
+    return {
+        ...config,
+        enabled: config.enabled ?? true,
+        _isAgent_: true,
+    }
+}
+
+/**
+ * Declares an agent tool — `define.Tool({ id, description, inputSchema, execute })`.
+ * The `id` is **required** (it is the name the model calls).
+ * A `define.McpTool` is accepted by an agent as-is.
+ */
+function Tool(tool: AgentToolDefinition): AgentToolDefinition {
+    return {
+        ...tool,
+        enabled: tool.enabled ?? true,
+        _isTool_: true,
+    }
+}
+
 
 
 export const define = {
@@ -186,5 +226,7 @@ export const define = {
   McpTool,
   McpResource,
   Lifecycle,
-  Vars
+  Vars,
+  Agent,
+  Tool,
 }

package/lib/endpoints.ts

@@ -0,0 +1,80 @@
+/**
+ * The URL of everything — one place.
+ *
+ * Every framework surface lives under a **single namespace**, `/api/:tenant_id`,
+ * and a **family** segment says what it is:
+ *
+ * | Surface | Route |
+ * | --- | --- |
+ * | Collections (CRUD, actions, login…) | `POST /api/:tenant_id/collections/:collection/:action` |
+ * | Services | `POST /api/:tenant_id/services/:service/:action` |
+ * | Variables | `POST /api/:tenant_id/vars/:action` |
+ * | Upload | `POST /api/:tenant_id/upload/:collection` |
+ * | Files (serve, delete) | `GET\|DELETE /api/:tenant_id/files/:collection/:file` |
+ * | Agents | `POST /api/:tenant_id/agents/:agent/:action` |
+ * | MCP (tools & resources) | `GET\|POST /api/:tenant_id/mcp` |
+ *
+ * Two reasons for the shape. A tenant's own routes (`routes: { prefix: '/api/v1' }`)
+ * are mounted **before** this one, so they always win — and nothing is ambiguous:
+ * a collection may be called `services` or `login` without ever colliding with the
+ * family segment, since a collection always sits under `/collections/`.
+ *
+ * `patterns` is what the server registers (Hono patterns); `endpoints` is what a
+ * client builds — the two are read together, so a path is never written twice.
+ */
+
+/** The family segments — the component right after the tenant id. */
+const FAMILIES = ["collections", "services", "vars", "upload", "files", "agents", "mcp"] as const;
+
+const patterns = {
+    collection: "/api/:tenant_id/collections/:collection/:action",
+    /** The login action of an auth-enabled collection — its own rate limit. */
+    login: "/api/:tenant_id/collections/:collection/login",
+    service: "/api/:tenant_id/services/:service/:action",
+    vars: "/api/:tenant_id/vars/:action",
+    upload: "/api/:tenant_id/upload/:collection",
+    file: "/api/:tenant_id/files/:collection/:file",
+    agent: "/api/:tenant_id/agents/:agent/:action",
+    mcp: "/api/:tenant_id/mcp",
+} as const;
+
+/** One path segment, encoded. */
+function segment(value: string): string {
+    return encodeURIComponent(String(value));
+}
+
+const endpoints = {
+    /** Everything of a collection — `find`, `insertOne`, `login`, a custom action… */
+    collection: (tenant: string, collection: string, action: string) =>
+        `/api/${segment(tenant)}/collections/${segment(collection)}/${segment(action)}`,
+
+    service: (tenant: string, service: string, action: string) =>
+        `/api/${segment(tenant)}/services/${segment(service)}/${segment(action)}`,
+
+    /** Variables — the namespace and the key travel in the body. */
+    vars: (tenant: string, action: string) =>
+        `/api/${segment(tenant)}/vars/${segment(action)}`,
+
+    upload: (tenant: string, collection: string) =>
+        `/api/${segment(tenant)}/upload/${segment(collection)}`,
+
+    /** One file of a file collection — the document id, or its filename. */
+    file: (tenant: string, collection: string, file: string) =>
+        `/api/${segment(tenant)}/files/${segment(collection)}/${segment(file)}`,
+
+    agent: (tenant: string, agent: string, action: string) =>
+        `/api/${segment(tenant)}/agents/${segment(agent)}/${segment(action)}`,
+
+    mcp: (tenant: string) => `/api/${segment(tenant)}/mcp`,
+} as const;
+
+/**
+ * Is this a segment the framework owns? A tenant route mounted at
+ * `/api/:tenant_id/<family>/…` is registered **first** and would shadow the
+ * framework API silently — the route loader warns about it.
+ */
+function isReservedFamily(name: string): boolean {
+    return (FAMILIES as readonly string[]).includes(name);
+}
+
+export { endpoints, FAMILIES, isReservedFamily, patterns, segment };

package/lib/files.ts

@@ -5,6 +5,7 @@ import { toBson } from "../utils/func";
 import { useRest } from "../database/rest";
 import { cfg } from "../server/config";
 import { AppError } from "./error";
+import { endpoints } from "./endpoints";
 import { S3Client } from "bun";
 import type { FileCollection } from "../types/file";
 import path from "path";
@@ -360,7 +361,7 @@ export async function handleUpload(options: UploadOptions): Promise<FileResult>
   const { path: filepath, size } = await storage.save(tenant_id, collection, file, { id: _id, mimetype, subpath });
 
   // Update MongoDB with actual file info
-  const url = `/files/${tenant_id}/${collection}/${filename}`;
+  const url = endpoints.file(tenant_id, collection, filename);
   try {
     const rest = new useRest({ tenant_id, internal: true, useHook: false, useCustomApi: false });
     await rest.db.collection(collection).updateOne(